Researchers at the University of Toronto say they have developed an algorithm that can learn directly from human instructions rather than a set of examples. Researchers usually provide neural networks with labeled data and teach the system how to make decisions based on the samples. With the new heuristic training model, humans program the algorithm with instructions that are used to classify training samples.
Researchers Parham Aarabi and Wenzhi Guo trained their algorithm to identify people's hair in photographs. "Our algorithm learned to correctly classify difficult, borderline cases — distinguishing the texture of hair versus the texture of the background," says Aarabi. "What we saw was like a teacher instructing a child, and the child learning beyond what the teacher taught her initially."
Using the new method, the algorithm outperformed conventional training techniques by 160 percent and outperformed its own training by 9 percent. The researchers say the heuristic method could be used to classify previously unlabeled data, such as cancerous cells for medical diagnostics or to classify objects surrounding a self-driving car.
From University of Toronto
View Full Article
Abstracts Copyright © 2016 Information Inc., Bethesda, Maryland, USA